Demo of "Modeling Interpretation Variations in Music Performance Rendering using CVRNN"
Akira Maezawa, Kazuhiko Yamamoto, Takuya Fujishima (Yamaha Corporation)
Correspondence: akira.maezawa __at__ music.yamaha.com
Demonstrations
(Loading the audio might take a few minutes)
Baroque
Classic
Romantic
Interpretation sequence sampled from the prior
Same as left, but the interpretation sequence has been offset. Notice that it generates a much calmer and shorter articulation.
Ragtime
Interpretation sequence sampled from the prior
Same song as the left, but adjusted the interpretation sequence the same way as above. Notice that here the playing becomes calmer and softer. This shows that the property of interpretation vector remains consistent between pieces.
Some comparisons with the baselines
Song 1
Original MIDI (Raw)
Rendered with Finale
Rendered with the proposed method
Song 2
Original MIDI (Raw)
Rendered with Finale
Rendered with the proposed method
Visualizing the piano-roll as interpretation vector is changed
Here, we can see that the articulation changes (as well as dynamics and the tempo) as we change the interpretation vector. The audio here plays the bottom two piano-rolls, first the left followed by the right. You can hear subtle nuances in the articulation as well as the tempo
Performance generated with the default interpretation vector sampled from the prior
Performance generated with the default interpretation vector sampled from the prior, with additional bias